China is shaking up the artificial intelligence industry with DeepSeek, a company that combines technological innovation with bold strategy. However, beyond the initial excitement, the "DeepSeek moment" appears to be as much a technical feat as a clever maneuver in an intense global competition.
A "DeepSeek Moment" Like Sputnik?
In 1957, Sputnik 1 triggered a space race between superpowers. In January 2025, DeepSeek seems to have replicated this effect in AI by launching two cutting-edge models: DeepSeek-V3 and DeepSeek-R1. These large language models (LLMs), trained on massive datasets, position themselves as serious competitors to OpenAI and Google, promising high performance at significantly reduced costs. However, this success has sparked both admiration and skepticism.
A Technological Breakthrough or an Economic Illusion?
DeepSeek's models stand out due to their open-source approach, in stark contrast to OpenAI or Google's proprietary choices. This transparency has been praised, notably by Yann Le Cun, who highlighted the power of collaborative innovation. But what truly caught attention was DeepSeek’s claim that it trained its V3 model for only $5.575 million. Compared to its competitors' astronomical investments, this figure seems almost impossible.
In reality, this amount covers only a fraction of the actual costs. The training of V3 was likely supported by a massive computing infrastructure accumulated by DeepSeek’s parent company, High-Flyer, long before the U.S. imposed export restrictions on advanced technologies to China. With a fleet of 50,000 GPUs, total expenses are estimated to be around $1.6 billion.
Three Innovations That Cut Costs
DeepSeek leverages impressive technical advances to optimize its models:
- Multi-Token Prediction (MTP): Anticipates multiple words at once, reducing computational effort.
- Mixture of Experts (MoE): Distributes processing across specialized "experts," reducing resource consumption.
- Multi-Head Latent Attention (MLA): Enhances memory management during interactions.
These innovations allow DeepSeek to reduce inference costs by up to 30 times compared to OpenAI, making AI more accessible to a broader audience.
Controversies and Suspicions
However, DeepSeek is surrounded by controversy. OpenAI has accused the company of using data distillation techniques to train its models on protected information. Additionally, some observers, like Palmer Luckey, suggest that DeepSeek's announcements are part of a psychological strategy aimed at destabilizing U.S. players and diverting attention from potential violations of technology sanctions.
A Lesson for Europe
Despite the controversies, DeepSeek serves as a reminder that innovation is still possible with limited resources. Europe, often lagging behind American and Chinese giants, could take inspiration from this approach. Companies like France’s Mistral demonstrate that agile teams and open-source strategies can compete with the largest organizations.
DeepSeek represents both a technological achievement and a well-orchestrated geopolitical strategy. While its true impact remains to be seen, its open-source model and technical innovations could redefine the global approach to artificial intelligence. However, the story of this Chinese AI also highlights the geopolitical and economic stakes that drive this relentless race toward the digital future.
